Qwen-Image (#12055)

* (feat): qwen-image integration

* fix(qwen-image):
- remove unused logics related to controlnet/ip-adapter

* fix(qwen-image):
- compatible with attention dispatcher
- cond cache support

* fix(qwen-image):
- cond cache registry
- attention backend argument
- fix copies

* fix(qwen-image):
- remove local test

* Update src/diffusers/models/transformers/transformer_qwenimage.py

---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>

src/diffusers/__init__.py

@@ -174,6 +174,7 @@ else:
             "AutoencoderKLLTXVideo",
             "AutoencoderKLMagvit",
             "AutoencoderKLMochi",
+            "AutoencoderKLQwenImage",
             "AutoencoderKLTemporalDecoder",
             "AutoencoderKLWan",
             "AutoencoderOobleck",
@@ -215,6 +216,7 @@ else:
             "OmniGenTransformer2DModel",
             "PixArtTransformer2DModel",
             "PriorTransformer",
+            "QwenImageTransformer2DModel",
             "SanaControlNetModel",
             "SanaTransformer2DModel",
             "SD3ControlNetModel",
@@ -486,6 +488,7 @@ else:
             "PixArtAlphaPipeline",
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
+            "QwenImagePipeline",
             "ReduxImageEncoder",
             "SanaControlNetPipeline",
             "SanaPAGPipeline",
@@ -832,6 +835,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLLTXVideo,
             AutoencoderKLMagvit,
             AutoencoderKLMochi,
+            AutoencoderKLQwenImage,
             AutoencoderKLTemporalDecoder,
             AutoencoderKLWan,
             AutoencoderOobleck,
@@ -873,6 +877,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             OmniGenTransformer2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            QwenImageTransformer2DModel,
             SanaControlNetModel,
             SanaTransformer2DModel,
             SD3ControlNetModel,
@@ -1119,6 +1124,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PixArtAlphaPipeline,
             PixArtSigmaPAGPipeline,
             PixArtSigmaPipeline,
+            QwenImagePipeline,
             ReduxImageEncoder,
             SanaControlNetPipeline,
             SanaPAGPipeline,

src/diffusers/hooks/_helpers.py

@@ -153,6 +153,7 @@ def _register_transformer_blocks_metadata():
     )
     from ..models.transformers.transformer_ltx import LTXVideoTransformerBlock
     from ..models.transformers.transformer_mochi import MochiTransformerBlock
+    from ..models.transformers.transformer_qwenimage import QwenImageTransformerBlock
     from ..models.transformers.transformer_wan import WanTransformerBlock
 
     # BasicTransformerBlock
@@ -255,6 +256,15 @@ def _register_transformer_blocks_metadata():
         ),
     )
 
+    # QwenImage
+    TransformerBlockRegistry.register(
+        model_class=QwenImageTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            return_hidden_states_index=1,
+            return_encoder_hidden_states_index=0,
+        ),
+    )
+
 
 # fmt: off
 def _skip_attention___ret___hidden_states(self, *args, **kwargs):

src/diffusers/models/__init__.py

@@ -38,6 +38,7 @@ if is_torch_available():
     _import_structure["autoencoders.autoencoder_kl_ltx"] = ["AutoencoderKLLTXVideo"]
     _import_structure["autoencoders.autoencoder_kl_magvit"] = ["AutoencoderKLMagvit"]
     _import_structure["autoencoders.autoencoder_kl_mochi"] = ["AutoencoderKLMochi"]
+    _import_structure["autoencoders.autoencoder_kl_qwenimage"] = ["AutoencoderKLQwenImage"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoders.autoencoder_kl_wan"] = ["AutoencoderKLWan"]
     _import_structure["autoencoders.autoencoder_oobleck"] = ["AutoencoderOobleck"]
@@ -88,6 +89,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_lumina2"] = ["Lumina2Transformer2DModel"]
     _import_structure["transformers.transformer_mochi"] = ["MochiTransformer3DModel"]
     _import_structure["transformers.transformer_omnigen"] = ["OmniGenTransformer2DModel"]
+    _import_structure["transformers.transformer_qwenimage"] = ["QwenImageTransformer2DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_skyreels_v2"] = ["SkyReelsV2Transformer3DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
@@ -126,6 +128,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLLTXVideo,
             AutoencoderKLMagvit,
             AutoencoderKLMochi,
+            AutoencoderKLQwenImage,
             AutoencoderKLTemporalDecoder,
             AutoencoderKLWan,
             AutoencoderOobleck,
@@ -177,6 +180,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             OmniGenTransformer2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            QwenImageTransformer2DModel,
             SanaTransformer2DModel,
             SD3Transformer2DModel,
             SkyReelsV2Transformer3DModel,

src/diffusers/models/autoencoders/__init__.py

@@ -8,6 +8,7 @@ from .autoencoder_kl_hunyuan_video import AutoencoderKLHunyuanVideo
 from .autoencoder_kl_ltx import AutoencoderKLLTXVideo
 from .autoencoder_kl_magvit import AutoencoderKLMagvit
 from .autoencoder_kl_mochi import AutoencoderKLMochi
+from .autoencoder_kl_qwenimage import AutoencoderKLQwenImage
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
 from .autoencoder_kl_wan import AutoencoderKLWan
 from .autoencoder_oobleck import AutoencoderOobleck

src/diffusers/models/autoencoders/autoencoder_kl_qwenimage.py

+# Copyright 2025 The Qwen-Image Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin
+from ...utils import logging
+from ...utils.accelerate_utils import apply_forward_hook
+from ..activations import get_activation
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from .vae import DecoderOutput, DiagonalGaussianDistribution
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+CACHE_T = 2
+
+
+class QwenImageCausalConv3d(nn.Conv3d):
+    r"""
+    A custom 3D causal convolution layer with feature caching support.
+
+    This layer extends the standard Conv3D layer by ensuring causality in the time dimension and handling feature
+    caching for efficient inference.
+
+    Args:
+        in_channels (int): Number of channels in the input image
+        out_channels (int): Number of channels produced by the convolution
+        kernel_size (int or tuple): Size of the convolving kernel
+        stride (int or tuple, optional): Stride of the convolution. Default: 1
+        padding (int or tuple, optional): Zero-padding added to all three sides of the input. Default: 0
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Union[int, Tuple[int, int, int]] = 1,
+        padding: Union[int, Tuple[int, int, int]] = 0,
+    ) -> None:
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+        )
+
+        # Set up causal padding
+        self._padding = (self.padding[2], self.padding[2], self.padding[1], self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+        return super().forward(x)
+
+
+class QwenImageRMS_norm(nn.Module):
+    r"""
+    A custom RMS normalization layer.
+
+    Args:
+        dim (int): The number of dimensions to normalize over.
+        channel_first (bool, optional): Whether the input tensor has channels as the first dimension.
+            Default is True.
+        images (bool, optional): Whether the input represents image data. Default is True.
+        bias (bool, optional): Whether to include a learnable bias term. Default is False.
+    """
+
+    def __init__(self, dim: int, channel_first: bool = True, images: bool = True, bias: bool = False) -> None:
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.0
+
+    def forward(self, x):
+        return F.normalize(x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma + self.bias
+
+
+class QwenImageUpsample(nn.Upsample):
+    r"""
+    Perform upsampling while ensuring the output tensor has the same data type as the input.
+
+    Args:
+        x (torch.Tensor): Input tensor to be upsampled.
+
+    Returns:
+        torch.Tensor: Upsampled tensor with the same data type as the input.
+    """
+
+    def forward(self, x):
+        return super().forward(x.float()).type_as(x)
+
+
+class QwenImageResample(nn.Module):
+    r"""
+    A custom resampling module for 2D and 3D data.
+
+    Args:
+        dim (int): The number of input/output channels.
+        mode (str): The resampling mode. Must be one of:
+            - 'none': No resampling (identity operation).
+            - 'upsample2d': 2D upsampling with nearest-exact interpolation and convolution.
+            - 'upsample3d': 3D upsampling with nearest-exact interpolation, convolution, and causal 3D convolution.
+            - 'downsample2d': 2D downsampling with zero-padding and convolution.
+            - 'downsample3d': 3D downsampling with zero-padding, convolution, and causal 3D convolution.
+    """
+
+    def __init__(self, dim: int, mode: str) -> None:
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                QwenImageUpsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim // 2, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                QwenImageUpsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim // 2, 3, padding=1),
+            )
+            self.time_conv = QwenImageCausalConv3d(dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = QwenImageCausalConv3d(dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == "upsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = "Rep"
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] != "Rep":
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat(
+                            [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2
+                        )
+                    if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] == "Rep":
+                        cache_x = torch.cat([torch.zeros_like(cache_x).to(cache_x.device), cache_x], dim=2)
+                    if feat_cache[idx] == "Rep":
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]), 3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.resample(x)
+        x = x.view(b, t, x.size(1), x.size(2), x.size(3)).permute(0, 2, 1, 3, 4)
+
+        if self.mode == "downsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+
+class QwenImageResidualBlock(nn.Module):
+    r"""
+    A custom residual block module.
+
+    Args:
+        in_dim (int): Number of input channels.
+        out_dim (int): Number of output channels.
+        dropout (float, optional): Dropout rate for the dropout layer. Default is 0.0.
+        non_linearity (str, optional): Type of non-linearity to use. Default is "silu".
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        dropout: float = 0.0,
+        non_linearity: str = "silu",
+    ) -> None:
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.nonlinearity = get_activation(non_linearity)
+
+        # layers
+        self.norm1 = QwenImageRMS_norm(in_dim, images=False)
+        self.conv1 = QwenImageCausalConv3d(in_dim, out_dim, 3, padding=1)
+        self.norm2 = QwenImageRMS_norm(out_dim, images=False)
+        self.dropout = nn.Dropout(dropout)
+        self.conv2 = QwenImageCausalConv3d(out_dim, out_dim, 3, padding=1)
+        self.conv_shortcut = QwenImageCausalConv3d(in_dim, out_dim, 1) if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        # Apply shortcut connection
+        h = self.conv_shortcut(x)
+
+        # First normalization and activation
+        x = self.norm1(x)
+        x = self.nonlinearity(x)
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        # Second normalization and activation
+        x = self.norm2(x)
+        x = self.nonlinearity(x)
+
+        # Dropout
+        x = self.dropout(x)
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
+            x = self.conv2(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv2(x)
+
+        # Add residual connection
+        return x + h
+
+
+class QwenImageAttentionBlock(nn.Module):
+    r"""
+    Causal self-attention with a single head.
+
+    Args:
+        dim (int): The number of channels in the input tensor.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = QwenImageRMS_norm(dim)
+        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
+        self.proj = nn.Conv2d(dim, dim, 1)
+
+    def forward(self, x):
+        identity = x
+        batch_size, channels, time, height, width = x.size()
+
+        x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * time, channels, height, width)
+        x = self.norm(x)
+
+        # compute query, key, value
+        qkv = self.to_qkv(x)
+        qkv = qkv.reshape(batch_size * time, 1, channels * 3, -1)
+        qkv = qkv.permute(0, 1, 3, 2).contiguous()
+        q, k, v = qkv.chunk(3, dim=-1)
+
+        # apply attention
+        x = F.scaled_dot_product_attention(q, k, v)
+
+        x = x.squeeze(1).permute(0, 2, 1).reshape(batch_size * time, channels, height, width)
+
+        # output projection
+        x = self.proj(x)
+
+        # Reshape back: [(b*t), c, h, w] -> [b, c, t, h, w]
+        x = x.view(batch_size, time, channels, height, width)
+        x = x.permute(0, 2, 1, 3, 4)
+
+        return x + identity
+
+
+class QwenImageMidBlock(nn.Module):
+    """
+    Middle block for QwenImageVAE encoder and decoder.
+
+    Args:
+        dim (int): Number of input/output channels.
+        dropout (float): Dropout rate.
+        non_linearity (str): Type of non-linearity to use.
+    """
+
+    def __init__(self, dim: int, dropout: float = 0.0, non_linearity: str = "silu", num_layers: int = 1):
+        super().__init__()
+        self.dim = dim
+
+        # Create the components
+        resnets = [QwenImageResidualBlock(dim, dim, dropout, non_linearity)]
+        attentions = []
+        for _ in range(num_layers):
+            attentions.append(QwenImageAttentionBlock(dim))
+            resnets.append(QwenImageResidualBlock(dim, dim, dropout, non_linearity))
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        # First residual block
+        x = self.resnets[0](x, feat_cache, feat_idx)
+
+        # Process through attention and residual blocks
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                x = attn(x)
+
+            x = resnet(x, feat_cache, feat_idx)
+
+        return x
+
+
+class QwenImageEncoder3d(nn.Module):
+    r"""
+    A 3D encoder module.
+
+    Args:
+        dim (int): The base number of channels in the first layer.
+        z_dim (int): The dimensionality of the latent space.
+        dim_mult (list of int): Multipliers for the number of channels in each block.
+        num_res_blocks (int): Number of residual blocks in each block.
+        attn_scales (list of float): Scales at which to apply attention mechanisms.
+        temperal_downsample (list of bool): Whether to downsample temporally in each block.
+        dropout (float): Dropout rate for the dropout layers.
+        non_linearity (str): Type of non-linearity to use.
+    """
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+        non_linearity: str = "silu",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.nonlinearity = get_activation(non_linearity)
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv_in = QwenImageCausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        self.down_blocks = nn.ModuleList([])
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                self.down_blocks.append(QwenImageResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    self.down_blocks.append(QwenImageAttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = "downsample3d" if temperal_downsample[i] else "downsample2d"
+                self.down_blocks.append(QwenImageResample(out_dim, mode=mode))
+                scale /= 2.0
+
+        # middle blocks
+        self.mid_block = QwenImageMidBlock(out_dim, dropout, non_linearity, num_layers=1)
+
+        # output blocks
+        self.norm_out = QwenImageRMS_norm(out_dim, images=False)
+        self.conv_out = QwenImageCausalConv3d(out_dim, z_dim, 3, padding=1)
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_in(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_in(x)
+
+        ## downsamples
+        for layer in self.down_blocks:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        x = self.mid_block(x, feat_cache, feat_idx)
+
+        ## head
+        x = self.norm_out(x)
+        x = self.nonlinearity(x)
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_out(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_out(x)
+        return x
+
+
+class QwenImageUpBlock(nn.Module):
+    """
+    A block that handles upsampling for the QwenImageVAE decoder.
+
+    Args:
+        in_dim (int): Input dimension
+        out_dim (int): Output dimension
+        num_res_blocks (int): Number of residual blocks
+        dropout (float): Dropout rate
+        upsample_mode (str, optional): Mode for upsampling ('upsample2d' or 'upsample3d')
+        non_linearity (str): Type of non-linearity to use
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        num_res_blocks: int,
+        dropout: float = 0.0,
+        upsample_mode: Optional[str] = None,
+        non_linearity: str = "silu",
+    ):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # Create layers list
+        resnets = []
+        # Add residual blocks and attention if needed
+        current_dim = in_dim
+        for _ in range(num_res_blocks + 1):
+            resnets.append(QwenImageResidualBlock(current_dim, out_dim, dropout, non_linearity))
+            current_dim = out_dim
+
+        self.resnets = nn.ModuleList(resnets)
+
+        # Add upsampling layer if needed
+        self.upsamplers = None
+        if upsample_mode is not None:
+            self.upsamplers = nn.ModuleList([QwenImageResample(out_dim, mode=upsample_mode)])
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        """
+        Forward pass through the upsampling block.
+
+        Args:
+            x (torch.Tensor): Input tensor
+            feat_cache (list, optional): Feature cache for causal convolutions
+            feat_idx (list, optional): Feature index for cache management
+
+        Returns:
+            torch.Tensor: Output tensor
+        """
+        for resnet in self.resnets:
+            if feat_cache is not None:
+                x = resnet(x, feat_cache, feat_idx)
+            else:
+                x = resnet(x)
+
+        if self.upsamplers is not None:
+            if feat_cache is not None:
+                x = self.upsamplers[0](x, feat_cache, feat_idx)
+            else:
+                x = self.upsamplers[0](x)
+        return x
+
+
+class QwenImageDecoder3d(nn.Module):
+    r"""
+    A 3D decoder module.
+
+    Args:
+        dim (int): The base number of channels in the first layer.
+        z_dim (int): The dimensionality of the latent space.
+        dim_mult (list of int): Multipliers for the number of channels in each block.
+        num_res_blocks (int): Number of residual blocks in each block.
+        attn_scales (list of float): Scales at which to apply attention mechanisms.
+        temperal_upsample (list of bool): Whether to upsample temporally in each block.
+        dropout (float): Dropout rate for the dropout layers.
+        non_linearity (str): Type of non-linearity to use.
+    """
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_upsample=[False, True, True],
+        dropout=0.0,
+        non_linearity: str = "silu",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        self.nonlinearity = get_activation(non_linearity)
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2 ** (len(dim_mult) - 2)
+
+        # init block
+        self.conv_in = QwenImageCausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.mid_block = QwenImageMidBlock(dims[0], dropout, non_linearity, num_layers=1)
+
+        # upsample blocks
+        self.up_blocks = nn.ModuleList([])
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i > 0:
+                in_dim = in_dim // 2
+
+            # Determine if we need upsampling
+            upsample_mode = None
+            if i != len(dim_mult) - 1:
+                upsample_mode = "upsample3d" if temperal_upsample[i] else "upsample2d"
+
+            # Create and add the upsampling block
+            up_block = QwenImageUpBlock(
+                in_dim=in_dim,
+                out_dim=out_dim,
+                num_res_blocks=num_res_blocks,
+                dropout=dropout,
+                upsample_mode=upsample_mode,
+                non_linearity=non_linearity,
+            )
+            self.up_blocks.append(up_block)
+
+            # Update scale for next iteration
+            if upsample_mode is not None:
+                scale *= 2.0
+
+        # output blocks
+        self.norm_out = QwenImageRMS_norm(out_dim, images=False)
+        self.conv_out = QwenImageCausalConv3d(out_dim, 3, 3, padding=1)
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_in(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_in(x)
+
+        ## middle
+        x = self.mid_block(x, feat_cache, feat_idx)
+
+        ## upsamples
+        for up_block in self.up_blocks:
+            x = up_block(x, feat_cache, feat_idx)
+
+        ## head
+        x = self.norm_out(x)
+        x = self.nonlinearity(x)
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_out(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_out(x)
+        return x
+
+
+class AutoencoderKLQwenImage(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    r"""
+    A VAE model with KL loss for encoding videos into latents and decoding latent representations into videos.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+    """
+
+    _supports_gradient_checkpointing = False
+
+    @register_to_config
+    def __init__(
+        self,
+        base_dim: int = 96,
+        z_dim: int = 16,
+        dim_mult: Tuple[int] = [1, 2, 4, 4],
+        num_res_blocks: int = 2,
+        attn_scales: List[float] = [],
+        temperal_downsample: List[bool] = [False, True, True],
+        dropout: float = 0.0,
+        latents_mean: List[float] = [
+            -0.7571,
+            -0.7089,
+            -0.9113,
+            0.1075,
+            -0.1745,
+            0.9653,
+            -0.1517,
+            1.5508,
+            0.4134,
+            -0.0715,
+            0.5517,
+            -0.3632,
+            -0.1922,
+            -0.9497,
+            0.2503,
+            -0.2921,
+        ],
+        latents_std: List[float] = [
+            2.8184,
+            1.4541,
+            2.3275,
+            2.6558,
+            1.2196,
+            1.7708,
+            2.6052,
+            2.0743,
+            3.2687,
+            2.1526,
+            2.8652,
+            1.5579,
+            1.6382,
+            1.1253,
+            2.8251,
+            1.9160,
+        ],
+    ) -> None:
+        super().__init__()
+
+        self.z_dim = z_dim
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        self.encoder = QwenImageEncoder3d(
+            base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout
+        )
+        self.quant_conv = QwenImageCausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.post_quant_conv = QwenImageCausalConv3d(z_dim, z_dim, 1)
+
+        self.decoder = QwenImageDecoder3d(
+            base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout
+        )
+
+        self.spatial_compression_ratio = 2 ** len(self.temperal_downsample)
+
+        # When decoding a batch of video latents at a time, one can save memory by slicing across the batch dimension
+        # to perform decoding of a single video latent at a time.
+        self.use_slicing = False
+
+        # When decoding spatially large video latents, the memory requirement is very high. By breaking the video latent
+        # frames spatially into smaller tiles and performing multiple forward passes for decoding, and then blending the
+        # intermediate tiles together, the memory requirement can be lowered.
+        self.use_tiling = False
+
+        # The minimal tile height and width for spatial tiling to be used
+        self.tile_sample_min_height = 256
+        self.tile_sample_min_width = 256
+
+        # The minimal distance between two spatial tiles
+        self.tile_sample_stride_height = 192
+        self.tile_sample_stride_width = 192
+
+        # Precompute and cache conv counts for encoder and decoder for clear_cache speedup
+        self._cached_conv_counts = {
+            "decoder": sum(isinstance(m, QwenImageCausalConv3d) for m in self.decoder.modules())
+            if self.decoder is not None
+            else 0,
+            "encoder": sum(isinstance(m, QwenImageCausalConv3d) for m in self.encoder.modules())
+            if self.encoder is not None
+            else 0,
+        }
+
+    def enable_tiling(
+        self,
+        tile_sample_min_height: Optional[int] = None,
+        tile_sample_min_width: Optional[int] = None,
+        tile_sample_stride_height: Optional[float] = None,
+        tile_sample_stride_width: Optional[float] = None,
+    ) -> None:
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+
+        Args:
+            tile_sample_min_height (`int`, *optional*):
+                The minimum height required for a sample to be separated into tiles across the height dimension.
+            tile_sample_min_width (`int`, *optional*):
+                The minimum width required for a sample to be separated into tiles across the width dimension.
+            tile_sample_stride_height (`int`, *optional*):
+                The minimum amount of overlap between two consecutive vertical tiles. This is to ensure that there are
+                no tiling artifacts produced across the height dimension.
+            tile_sample_stride_width (`int`, *optional*):
+                The stride between two consecutive horizontal tiles. This is to ensure that there are no tiling
+                artifacts produced across the width dimension.
+        """
+        self.use_tiling = True
+        self.tile_sample_min_height = tile_sample_min_height or self.tile_sample_min_height
+        self.tile_sample_min_width = tile_sample_min_width or self.tile_sample_min_width
+        self.tile_sample_stride_height = tile_sample_stride_height or self.tile_sample_stride_height
+        self.tile_sample_stride_width = tile_sample_stride_width or self.tile_sample_stride_width
+
+    def disable_tiling(self) -> None:
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_tiling = False
+
+    def enable_slicing(self) -> None:
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self) -> None:
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    def clear_cache(self):
+        def _count_conv3d(model):
+            count = 0
+            for m in model.modules():
+                if isinstance(m, QwenImageCausalConv3d):
+                    count += 1
+            return count
+
+        self._conv_num = _count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = _count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num
+
+    def _encode(self, x: torch.Tensor):
+        _, _, num_frame, height, width = x.shape
+
+        if self.use_tiling and (width > self.tile_sample_min_width or height > self.tile_sample_min_height):
+            return self.tiled_encode(x)
+
+        self.clear_cache()
+        iter_ = 1 + (num_frame - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :], feat_cache=self._enc_feat_map, feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1) : 1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+
+        enc = self.quant_conv(out)
+        self.clear_cache()
+        return enc
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.Tensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        r"""
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded videos. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self._encode(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self._encode(x)
+        posterior = DiagonalGaussianDistribution(h)
+
+        if not return_dict:
+            return (posterior,)
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.Tensor, return_dict: bool = True):
+        _, _, num_frame, height, width = z.shape
+        tile_latent_min_height = self.tile_sample_min_height // self.spatial_compression_ratio
+        tile_latent_min_width = self.tile_sample_min_width // self.spatial_compression_ratio
+
+        if self.use_tiling and (width > tile_latent_min_width or height > tile_latent_min_height):
+            return self.tiled_decode(z, return_dict=return_dict)
+
+        self.clear_cache()
+        x = self.post_quant_conv(z)
+        for i in range(num_frame):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i : i + 1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(x[:, :, i : i + 1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+
+        out = torch.clamp(out, min=-1.0, max=1.0)
+        self.clear_cache()
+        if not return_dict:
+            return (out,)
+
+        return DecoderOutput(sample=out)
+
+    @apply_forward_hook
+    def decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z).sample
+
+        if not return_dict:
+            return (decoded,)
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[-2], b.shape[-2], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, :, y, :] = a[:, :, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, :, y, :] * (
+                y / blend_extent
+            )
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[-1], b.shape[-1], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, :, x] = a[:, :, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, :, x] * (
+                x / blend_extent
+            )
+        return b
+
+    def tiled_encode(self, x: torch.Tensor) -> AutoencoderKLOutput:
+        r"""Encode a batch of images using a tiled encoder.
+
+        Args:
+            x (`torch.Tensor`): Input batch of videos.
+
+        Returns:
+            `torch.Tensor`:
+                The latent representation of the encoded videos.
+        """
+        _, _, num_frames, height, width = x.shape
+        latent_height = height // self.spatial_compression_ratio
+        latent_width = width // self.spatial_compression_ratio
+
+        tile_latent_min_height = self.tile_sample_min_height // self.spatial_compression_ratio
+        tile_latent_min_width = self.tile_sample_min_width // self.spatial_compression_ratio
+        tile_latent_stride_height = self.tile_sample_stride_height // self.spatial_compression_ratio
+        tile_latent_stride_width = self.tile_sample_stride_width // self.spatial_compression_ratio
+
+        blend_height = tile_latent_min_height - tile_latent_stride_height
+        blend_width = tile_latent_min_width - tile_latent_stride_width
+
+        # Split x into overlapping tiles and encode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, height, self.tile_sample_stride_height):
+            row = []
+            for j in range(0, width, self.tile_sample_stride_width):
+                self.clear_cache()
+                time = []
+                frame_range = 1 + (num_frames - 1) // 4
+                for k in range(frame_range):
+                    self._enc_conv_idx = [0]
+                    if k == 0:
+                        tile = x[:, :, :1, i : i + self.tile_sample_min_height, j : j + self.tile_sample_min_width]
+                    else:
+                        tile = x[
+                            :,
+                            :,
+                            1 + 4 * (k - 1) : 1 + 4 * k,
+                            i : i + self.tile_sample_min_height,
+                            j : j + self.tile_sample_min_width,
+                        ]
+                    tile = self.encoder(tile, feat_cache=self._enc_feat_map, feat_idx=self._enc_conv_idx)
+                    tile = self.quant_conv(tile)
+                    time.append(tile)
+                row.append(torch.cat(time, dim=2))
+            rows.append(row)
+        self.clear_cache()
+
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_height)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_width)
+                result_row.append(tile[:, :, :, :tile_latent_stride_height, :tile_latent_stride_width])
+            result_rows.append(torch.cat(result_row, dim=-1))
+
+        enc = torch.cat(result_rows, dim=3)[:, :, :, :latent_height, :latent_width]
+        return enc
+
+    def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images using a tiled decoder.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        _, _, num_frames, height, width = z.shape
+        sample_height = height * self.spatial_compression_ratio
+        sample_width = width * self.spatial_compression_ratio
+
+        tile_latent_min_height = self.tile_sample_min_height // self.spatial_compression_ratio
+        tile_latent_min_width = self.tile_sample_min_width // self.spatial_compression_ratio
+        tile_latent_stride_height = self.tile_sample_stride_height // self.spatial_compression_ratio
+        tile_latent_stride_width = self.tile_sample_stride_width // self.spatial_compression_ratio
+
+        blend_height = self.tile_sample_min_height - self.tile_sample_stride_height
+        blend_width = self.tile_sample_min_width - self.tile_sample_stride_width
+
+        # Split z into overlapping tiles and decode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, height, tile_latent_stride_height):
+            row = []
+            for j in range(0, width, tile_latent_stride_width):
+                self.clear_cache()
+                time = []
+                for k in range(num_frames):
+                    self._conv_idx = [0]
+                    tile = z[:, :, k : k + 1, i : i + tile_latent_min_height, j : j + tile_latent_min_width]
+                    tile = self.post_quant_conv(tile)
+                    decoded = self.decoder(tile, feat_cache=self._feat_map, feat_idx=self._conv_idx)
+                    time.append(decoded)
+                row.append(torch.cat(time, dim=2))
+            rows.append(row)
+        self.clear_cache()
+
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_height)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_width)
+                result_row.append(tile[:, :, :, : self.tile_sample_stride_height, : self.tile_sample_stride_width])
+            result_rows.append(torch.cat(result_row, dim=-1))
+
+        dec = torch.cat(result_rows, dim=3)[:, :, :, :sample_height, :sample_width]
+
+        if not return_dict:
+            return (dec,)
+        return DecoderOutput(sample=dec)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput, torch.Tensor]:
+        """
+        Args:
+            sample (`torch.Tensor`): Input sample.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, return_dict=return_dict)
+        return dec

src/diffusers/models/transformers/__init__.py

@@ -30,6 +30,7 @@ if is_torch_available():
     from .transformer_lumina2 import Lumina2Transformer2DModel
     from .transformer_mochi import MochiTransformer3DModel
     from .transformer_omnigen import OmniGenTransformer2DModel
+    from .transformer_qwenimage import QwenImageTransformer2DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_skyreels_v2 import SkyReelsV2Transformer3DModel
     from .transformer_temporal import TransformerTemporalModel

src/diffusers/models/transformers/transformer_qwenimage.py

+# Copyright 2025 Qwen-Image Team, The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import FeedForward
+from ..attention_dispatch import dispatch_attention_fn
+from ..attention_processor import Attention
+from ..cache_utils import CacheMixin
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormContinuous, RMSNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_timestep_embedding(
+    timesteps: torch.Tensor,
+    embedding_dim: int,
+    flip_sin_to_cos: bool = False,
+    downscale_freq_shift: float = 1,
+    scale: float = 1,
+    max_period: int = 10000,
+) -> torch.Tensor:
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
+
+    Args
+        timesteps (torch.Tensor):
+            a 1-D Tensor of N indices, one per batch element. These may be fractional.
+        embedding_dim (int):
+            the dimension of the output.
+        flip_sin_to_cos (bool):
+            Whether the embedding order should be `cos, sin` (if True) or `sin, cos` (if False)
+        downscale_freq_shift (float):
+            Controls the delta between frequencies between dimensions
+        scale (float):
+            Scaling factor applied to the embeddings.
+        max_period (int):
+            Controls the maximum frequency of the embeddings
+    Returns
+        torch.Tensor: an [N x dim] Tensor of positional embeddings.
+    """
+    assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
+
+    half_dim = embedding_dim // 2
+    exponent = -math.log(max_period) * torch.arange(
+        start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
+    )
+    exponent = exponent / (half_dim - downscale_freq_shift)
+
+    emb = torch.exp(exponent).to(timesteps.dtype)
+    emb = timesteps[:, None].float() * emb[None, :]
+
+    # scale embeddings
+    emb = scale * emb
+
+    # concat sine and cosine embeddings
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+
+    # flip sine and cosine embeddings
+    if flip_sin_to_cos:
+        emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
+
+    # zero pad
+    if embedding_dim % 2 == 1:
+        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
+    return emb
+
+
+def apply_rotary_emb_qwen(
+    x: torch.Tensor,
+    freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+    use_real: bool = True,
+    use_real_unbind_dim: int = -1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+    to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+    reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+    tensors contain rotary embeddings and are returned as real tensors.
+
+    Args:
+        x (`torch.Tensor`):
+            Query or key tensor to apply rotary embeddings. [B, S, H, D] xk (torch.Tensor): Key tensor to apply
+        freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+    """
+    if use_real:
+        cos, sin = freqs_cis  # [S, D]
+        cos = cos[None, None]
+        sin = sin[None, None]
+        cos, sin = cos.to(x.device), sin.to(x.device)
+
+        if use_real_unbind_dim == -1:
+            # Used for flux, cogvideox, hunyuan-dit
+            x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+            x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+        elif use_real_unbind_dim == -2:
+            # Used for Stable Audio, OmniGen, CogView4 and Cosmos
+            x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2)  # [B, S, H, D//2]
+            x_rotated = torch.cat([-x_imag, x_real], dim=-1)
+        else:
+            raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.")
+
+        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+        return out
+    else:
+        x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
+        freqs_cis = freqs_cis.unsqueeze(1)
+        x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3)
+
+        return x_out.type_as(x)
+
+
+class QwenTimestepProjEmbeddings(nn.Module):
+    def __init__(self, embedding_dim, pooled_projection_dim):
+        super().__init__()
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+    def forward(self, timestep, hidden_states):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))  # (N, D)
+
+        conditioning = timesteps_emb
+
+        return conditioning
+
+
+class QwenEmbedRope(nn.Module):
+    def __init__(self, theta: int, axes_dim: List[int], scale_rope=False):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        pos_index = torch.arange(1024)
+        neg_index = torch.arange(1024).flip(0) * -1 - 1
+        self.pos_freqs = torch.cat(
+            [
+                self.rope_params(pos_index, self.axes_dim[0], self.theta),
+                self.rope_params(pos_index, self.axes_dim[1], self.theta),
+                self.rope_params(pos_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+        self.neg_freqs = torch.cat(
+            [
+                self.rope_params(neg_index, self.axes_dim[0], self.theta),
+                self.rope_params(neg_index, self.axes_dim[1], self.theta),
+                self.rope_params(neg_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+        self.rope_cache = {}
+
+        # 是否使用 scale rope
+        self.scale_rope = scale_rope
+
+    def rope_params(self, index, dim, theta=10000):
+        """
+        Args:
+            index: [0, 1, 2, 3] 1D Tensor representing the position index of the token
+        """
+        assert dim % 2 == 0
+        freqs = torch.outer(index, 1.0 / torch.pow(theta, torch.arange(0, dim, 2).to(torch.float32).div(dim)))
+        freqs = torch.polar(torch.ones_like(freqs), freqs)
+        return freqs
+
+    def forward(self, video_fhw, txt_seq_lens, device):
+        """
+        Args: video_fhw: [frame, height, width] a list of 3 integers representing the shape of the video Args:
+        txt_length: [bs] a list of 1 integers representing the length of the text
+        """
+        if self.pos_freqs.device != device:
+            self.pos_freqs = self.pos_freqs.to(device)
+            self.neg_freqs = self.neg_freqs.to(device)
+
+        if isinstance(video_fhw, list):
+            video_fhw = video_fhw[0]
+        frame, height, width = video_fhw
+        rope_key = f"{frame}_{height}_{width}"
+
+        if rope_key not in self.rope_cache:
+            seq_lens = frame * height * width
+            freqs_pos = self.pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+            freqs_neg = self.neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+            freqs_frame = freqs_pos[0][:frame].view(frame, 1, 1, -1).expand(frame, height, width, -1)
+            if self.scale_rope:
+                freqs_height = torch.cat([freqs_neg[1][-(height - height // 2) :], freqs_pos[1][: height // 2]], dim=0)
+                freqs_height = freqs_height.view(1, height, 1, -1).expand(frame, height, width, -1)
+                freqs_width = torch.cat([freqs_neg[2][-(width - width // 2) :], freqs_pos[2][: width // 2]], dim=0)
+                freqs_width = freqs_width.view(1, 1, width, -1).expand(frame, height, width, -1)
+
+            else:
+                freqs_height = freqs_pos[1][:height].view(1, height, 1, -1).expand(frame, height, width, -1)
+                freqs_width = freqs_pos[2][:width].view(1, 1, width, -1).expand(frame, height, width, -1)
+
+            freqs = torch.cat([freqs_frame, freqs_height, freqs_width], dim=-1).reshape(seq_lens, -1)
+            self.rope_cache[rope_key] = freqs.clone().contiguous()
+        vid_freqs = self.rope_cache[rope_key]
+
+        if self.scale_rope:
+            max_vid_index = max(height // 2, width // 2)
+        else:
+            max_vid_index = max(height, width)
+
+        max_len = max(txt_seq_lens)
+        txt_freqs = self.pos_freqs[max_vid_index : max_vid_index + max_len, ...]
+
+        return vid_freqs, txt_freqs
+
+
+class QwenDoubleStreamAttnProcessor2_0:
+    """
+    Attention processor for Qwen double-stream architecture, matching DoubleStreamLayerMegatron logic. This processor
+    implements joint attention computation where text and image streams are processed together.
+    """
+
+    _attention_backend = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "QwenDoubleStreamAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,  # Image stream
+        encoder_hidden_states: torch.FloatTensor = None,  # Text stream
+        encoder_hidden_states_mask: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        if encoder_hidden_states is None:
+            raise ValueError("QwenDoubleStreamAttnProcessor2_0 requires encoder_hidden_states (text stream)")
+
+        seq_txt = encoder_hidden_states.shape[1]
+
+        # Compute QKV for image stream (sample projections)
+        img_query = attn.to_q(hidden_states)
+        img_key = attn.to_k(hidden_states)
+        img_value = attn.to_v(hidden_states)
+
+        # Compute QKV for text stream (context projections)
+        txt_query = attn.add_q_proj(encoder_hidden_states)
+        txt_key = attn.add_k_proj(encoder_hidden_states)
+        txt_value = attn.add_v_proj(encoder_hidden_states)
+
+        # Reshape for multi-head attention
+        img_query = img_query.unflatten(-1, (attn.heads, -1))
+        img_key = img_key.unflatten(-1, (attn.heads, -1))
+        img_value = img_value.unflatten(-1, (attn.heads, -1))
+
+        txt_query = txt_query.unflatten(-1, (attn.heads, -1))
+        txt_key = txt_key.unflatten(-1, (attn.heads, -1))
+        txt_value = txt_value.unflatten(-1, (attn.heads, -1))
+
+        # Apply QK normalization
+        if attn.norm_q is not None:
+            img_query = attn.norm_q(img_query)
+        if attn.norm_k is not None:
+            img_key = attn.norm_k(img_key)
+        if attn.norm_added_q is not None:
+            txt_query = attn.norm_added_q(txt_query)
+        if attn.norm_added_k is not None:
+            txt_key = attn.norm_added_k(txt_key)
+
+        # Apply RoPE
+        if image_rotary_emb is not None:
+            img_freqs, txt_freqs = image_rotary_emb
+            img_query = apply_rotary_emb_qwen(img_query, img_freqs, use_real=False)
+            img_key = apply_rotary_emb_qwen(img_key, img_freqs, use_real=False)
+            txt_query = apply_rotary_emb_qwen(txt_query, txt_freqs, use_real=False)
+            txt_key = apply_rotary_emb_qwen(txt_key, txt_freqs, use_real=False)
+
+        # Concatenate for joint attention
+        # Order: [text, image]
+        joint_query = torch.cat([txt_query, img_query], dim=1)
+        joint_key = torch.cat([txt_key, img_key], dim=1)
+        joint_value = torch.cat([txt_value, img_value], dim=1)
+
+        # Compute joint attention
+        joint_hidden_states = dispatch_attention_fn(
+            joint_query,
+            joint_key,
+            joint_value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+        )
+
+        # Reshape back
+        joint_hidden_states = joint_hidden_states.flatten(2, 3)
+        joint_hidden_states = joint_hidden_states.to(joint_query.dtype)
+
+        # Split attention outputs back
+        txt_attn_output = joint_hidden_states[:, :seq_txt, :]  # Text part
+        img_attn_output = joint_hidden_states[:, seq_txt:, :]  # Image part
+
+        # Apply output projections
+        img_attn_output = attn.to_out[0](img_attn_output)
+        if len(attn.to_out) > 1:
+            img_attn_output = attn.to_out[1](img_attn_output)  # dropout
+
+        txt_attn_output = attn.to_add_out(txt_attn_output)
+
+        return img_attn_output, txt_attn_output
+
+
+@maybe_allow_in_graph
+class QwenImageTransformerBlock(nn.Module):
+    def __init__(
+        self, dim: int, num_attention_heads: int, attention_head_dim: int, qk_norm: str = "rms_norm", eps: float = 1e-6
+    ):
+        super().__init__()
+
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+
+        # Image processing modules
+        self.img_mod = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(dim, 6 * dim, bias=True),  # For scale, shift, gate for norm1 and norm2
+        )
+        self.img_norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,  # Enable cross attention for joint computation
+            added_kv_proj_dim=dim,  # Enable added KV projections for text stream
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            context_pre_only=False,
+            bias=True,
+            processor=QwenDoubleStreamAttnProcessor2_0(),
+            qk_norm=qk_norm,
+            eps=eps,
+        )
+        self.img_norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.img_mlp = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+        # Text processing modules
+        self.txt_mod = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(dim, 6 * dim, bias=True),  # For scale, shift, gate for norm1 and norm2
+        )
+        self.txt_norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        # Text doesn't need separate attention - it's handled by img_attn joint computation
+        self.txt_norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.txt_mlp = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+    def _modulate(self, x, mod_params):
+        """Apply modulation to input tensor"""
+        shift, scale, gate = mod_params.chunk(3, dim=-1)
+        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1), gate.unsqueeze(1)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_mask: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Get modulation parameters for both streams
+        img_mod_params = self.img_mod(temb)  # [B, 6*dim]
+        txt_mod_params = self.txt_mod(temb)  # [B, 6*dim]
+
+        # Split modulation parameters for norm1 and norm2
+        img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)  # Each [B, 3*dim]
+        txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)  # Each [B, 3*dim]
+
+        # Process image stream - norm1 + modulation
+        img_normed = self.img_norm1(hidden_states)
+        img_modulated, img_gate1 = self._modulate(img_normed, img_mod1)
+
+        # Process text stream - norm1 + modulation
+        txt_normed = self.txt_norm1(encoder_hidden_states)
+        txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1)
+
+        # Use QwenAttnProcessor2_0 for joint attention computation
+        # This directly implements the DoubleStreamLayerMegatron logic:
+        # 1. Computes QKV for both streams
+        # 2. Applies QK normalization and RoPE
+        # 3. Concatenates and runs joint attention
+        # 4. Splits results back to separate streams
+        joint_attention_kwargs = joint_attention_kwargs or {}
+        attn_output = self.attn(
+            hidden_states=img_modulated,  # Image stream (will be processed as "sample")
+            encoder_hidden_states=txt_modulated,  # Text stream (will be processed as "context")
+            encoder_hidden_states_mask=encoder_hidden_states_mask,
+            image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
+        )
+
+        # QwenAttnProcessor2_0 returns (img_output, txt_output) when encoder_hidden_states is provided
+        img_attn_output, txt_attn_output = attn_output
+
+        # Apply attention gates and add residual (like in Megatron)
+        hidden_states = hidden_states + img_gate1 * img_attn_output
+        encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
+
+        # Process image stream - norm2 + MLP
+        img_normed2 = self.img_norm2(hidden_states)
+        img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
+        img_mlp_output = self.img_mlp(img_modulated2)
+        hidden_states = hidden_states + img_gate2 * img_mlp_output
+
+        # Process text stream - norm2 + MLP
+        txt_normed2 = self.txt_norm2(encoder_hidden_states)
+        txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
+        txt_mlp_output = self.txt_mlp(txt_modulated2)
+        encoder_hidden_states = encoder_hidden_states + txt_gate2 * txt_mlp_output
+
+        # Clip to prevent overflow for fp16
+        if encoder_hidden_states.dtype == torch.float16:
+            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+        if hidden_states.dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+
+        return encoder_hidden_states, hidden_states
+
+
+class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    """
+    The Transformer model introduced in Qwen.
+
+    Args:
+        patch_size (`int`, defaults to `2`):
+            Patch size to turn the input data into small patches.
+        in_channels (`int`, defaults to `64`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `None`):
+            The number of channels in the output. If not specified, it defaults to `in_channels`.
+        num_layers (`int`, defaults to `60`):
+            The number of layers of dual stream DiT blocks to use.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of dimensions to use for each attention head.
+        num_attention_heads (`int`, defaults to `24`):
+            The number of attention heads to use.
+        joint_attention_dim (`int`, defaults to `3584`):
+            The number of dimensions to use for the joint attention (embedding/channel dimension of
+            `encoder_hidden_states`).
+        pooled_projection_dim (`int`, defaults to `768`):
+            The number of dimensions to use for the pooled projection.
+        guidance_embeds (`bool`, defaults to `False`):
+            Whether to use guidance embeddings for guidance-distilled variant of the model.
+        axes_dims_rope (`Tuple[int]`, defaults to `(16, 56, 56)`):
+            The dimensions to use for the rotary positional embeddings.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["QwenImageTransformerBlock"]
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 64,
+        out_channels: Optional[int] = 16,
+        num_layers: int = 60,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        joint_attention_dim: int = 3584,
+        pooled_projection_dim: int = 768,
+        guidance_embeds: bool = False,
+        axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
+    ):
+        super().__init__()
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
+
+        self.time_text_embed = QwenTimestepProjEmbeddings(
+            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
+        )
+
+        self.txt_norm = RMSNorm(joint_attention_dim, eps=1e-6)
+
+        self.img_in = nn.Linear(in_channels, self.inner_dim)
+        self.txt_in = nn.Linear(joint_attention_dim, self.inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                QwenImageTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        encoder_hidden_states_mask: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        img_shapes: Optional[List[Tuple[int, int, int]]] = None,
+        txt_seq_lens: Optional[List[int]] = None,
+        guidance: torch.Tensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+        controlnet_blocks_repeat: bool = False,
+    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
+        """
+        The [`QwenTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            encoder_hidden_states_mask (`torch.Tensor` of shape `(batch_size, text_sequence_length)`):
+                Mask of the input conditions.
+            timestep ( `torch.LongTensor`):
+                Used to indicate denoising step.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        hidden_states = self.img_in(hidden_states)
+
+        timestep = timestep.to(hidden_states.dtype)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance.to(hidden_states.dtype) * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
+
+        image_rotary_emb = self.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device)
+
+        for index_block, block in enumerate(self.transformer_blocks):
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    encoder_hidden_states_mask,
+                    temb,
+                    image_rotary_emb,
+                )
+
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+
+        # Use only the image part (hidden_states) from the dual-stream blocks
+        hidden_states = self.norm_out(hidden_states, temb)
+        output = self.proj_out(hidden_states)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/pipelines/__init__.py

@@ -387,6 +387,7 @@ else:
         "SkyReelsV2ImageToVideoPipeline",
         "SkyReelsV2Pipeline",
     ]
+    _import_structure["qwenimage"] = ["QwenImagePipeline"]
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -703,6 +704,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
+        from .qwenimage import QwenImagePipeline
         from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintImg2ImgPipeline, SanaSprintPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline

src/diffusers/pipelines/qwenimage/__init__.py

+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_additional_imports = {}
+_import_structure = {"pipeline_output": ["QwenImagePipelineOutput", "QwenImagePriorReduxPipelineOutput"]}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["modeling_qwenimage"] = ["ReduxImageEncoder"]
+    _import_structure["pipeline_qwenimage"] = ["QwenImagePipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .pipeline_qwenimage import QwenImagePipeline
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)
+    for name, value in _additional_imports.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/qwenimage/pipeline_output.py

+from dataclasses import dataclass
+from typing import List, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import BaseOutput
+
+
+@dataclass
+class QwenImagePipelineOutput(BaseOutput):
+    """
+    Output class for Stable Diffusion pipelines.
+
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+    """
+
+    images: Union[List[PIL.Image.Image], np.ndarray]

src/diffusers/pipelines/qwenimage/pipeline_qwenimage.py

+# Copyright 2025 Qwen-Image Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import (
+    Qwen2_5_VLForConditionalGeneration,
+    Qwen2Tokenizer,
+)
+
+from ...image_processor import VaeImageProcessor
+from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import (
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import QwenImagePipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import QwenImagePipeline
+
+        >>> pipe = QwenImagePipeline.from_pretrained("Qwen/QwenImage-20B", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+        >>> prompt = "A cat holding a sign that says hello world"
+        >>> # Depending on the variant being used, the pipeline call will slightly vary.
+        >>> # Refer to the pipeline documentation for more details.
+        >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0]
+        >>> image.save("qwenimage.png")
+        ```
+"""
+
+
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class QwenImagePipeline(
+    DiffusionPipeline,
+):
+    r"""
+    The QwenImage pipeline for text-to-image generation.
+
+    Args:
+        transformer ([`QwenImageTransformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`Qwen2.5-VL-7B-Instruct`]):
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant.
+        tokenizer (`QwenTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = ["image_encoder", "feature_extractor"]
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLQwenImage,
+        text_encoder: Qwen2_5_VLForConditionalGeneration,
+        tokenizer: Qwen2Tokenizer,
+        transformer: QwenImageTransformer2DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
+        # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+        self.tokenizer_max_length = 1024
+        self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.prompt_template_encode_start_idx = 34
+        self.default_sample_size = 128
+
+    def extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor):
+        bool_mask = mask.bool()
+
+        valid_lengths = bool_mask.sum(dim=1)
+
+        selected = hidden_states[bool_mask]
+
+        split_result = torch.split(selected, valid_lengths.tolist(), dim=0)
+
+        return split_result
+
+    def _get_qwen_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_images_per_prompt: int = 1,
+        max_sequence_length: int = 1024,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        template = self.prompt_template_encode
+        drop_idx = self.prompt_template_encode_start_idx
+        txt = [template.format(e) for e in prompt]
+        txt_tokens = self.tokenizer(
+            txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt"
+        ).to(self.device)
+        encoder_hidden_states = self.text_encoder(
+            input_ids=txt_tokens.input_ids,
+            attention_mask=txt_tokens.attention_mask,
+            output_hidden_states=True,
+        )
+        hidden_states = encoder_hidden_states.hidden_states[-1]
+        split_hidden_states = self.extract_masked_hidden(hidden_states, txt_tokens.attention_mask)
+        split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
+        attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
+        max_seq_len = max([e.size(0) for e in split_hidden_states])
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states]
+        )
+        encoder_attention_mask = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list]
+        )
+
+        dtype = self.text_encoder.dtype
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        encoder_attention_mask = encoder_attention_mask.repeat(1, num_images_per_prompt, 1)
+        encoder_attention_mask = encoder_attention_mask.view(batch_size * num_images_per_prompt, seq_len)
+
+        return prompt_embeds, encoder_attention_mask
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_embeds_mask: Optional[torch.FloatTensor] = None,
+        max_sequence_length: int = 1024,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in all text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+            )
+
+        dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype
+        text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
+
+        return prompt_embeds, prompt_embeds_mask, text_ids
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_embeds_mask=None,
+        negative_prompt_embeds_mask=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
+            logger.warning(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_embeds_mask is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+        if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 1024:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}")
+
+    @staticmethod
+    def _prepare_latent_image_ids(batch_size, height, width, device, dtype):
+        latent_image_ids = torch.zeros(height, width, 3)
+        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None]
+        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :]
+
+        latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
+
+        latent_image_ids = latent_image_ids.reshape(
+            latent_image_id_height * latent_image_id_width, latent_image_id_channels
+        )
+
+        return latent_image_ids.to(device=device, dtype=dtype)
+
+    @staticmethod
+    def _pack_latents(latents, batch_size, num_channels_latents, height, width):
+        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
+        latents = latents.permute(0, 2, 4, 1, 3, 5)
+        latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)
+
+        return latents
+
+    @staticmethod
+    def _unpack_latents(latents, height, width, vae_scale_factor):
+        batch_size, num_patches, channels = latents.shape
+
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (vae_scale_factor * 2))
+        width = 2 * (int(width) // (vae_scale_factor * 2))
+
+        latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2)
+        latents = latents.permute(0, 3, 1, 4, 2, 5)
+
+        latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width)
+
+        return latents
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+
+        shape = (batch_size, 1, num_channels_latents, height, width)
+
+        if latents is not None:
+            latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
+            return latents.to(device=device, dtype=dtype), latent_image_ids
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
+
+        latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
+
+        return latents, latent_image_ids
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        true_cfg_scale: float = 4.0,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: float = 1.0,
+        num_images_per_prompt: int = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_embeds_mask: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds_mask: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is
+                not greater than `1`).
+            true_cfg_scale (`float`, *optional*, defaults to 1.0):
+                When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 3.5):
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
+                `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
+                the text `prompt`, usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will be generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
+            negative_ip_adapter_image:
+                (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            negative_ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is a list with the generated images.
+        """
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            negative_prompt_embeds_mask=negative_prompt_embeds_mask,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        has_neg_prompt = negative_prompt is not None or (
+            negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None
+        )
+        do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
+        (
+            prompt_embeds,
+            prompt_embeds_mask,
+            text_ids,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_embeds=prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+        )
+        if do_true_cfg:
+            (
+                negative_prompt_embeds,
+                negative_prompt_embeds_mask,
+                negative_text_ids,
+            ) = self.encode_prompt(
+                prompt=negative_prompt,
+                prompt_embeds=negative_prompt_embeds,
+                prompt_embeds_mask=negative_prompt_embeds_mask,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, latent_image_ids = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+        img_shapes = [(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)] * batch_size
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
+        image_seq_len = latents.shape[1]
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.15),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # print(f"timesteps: {timesteps}")
+
+        # handle guidance
+        if self.transformer.config.guidance_embeds:
+            guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
+            guidance = guidance.expand(latents.shape[0])
+        else:
+            guidance = None
+
+        if self.joint_attention_kwargs is None:
+            self._joint_attention_kwargs = {}
+
+        # 6. Denoising loop
+        self.scheduler.set_begin_index(0)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+                with self.transformer.cache_context("cond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latents,
+                        timestep=timestep / 1000,
+                        guidance=guidance,
+                        encoder_hidden_states_mask=prompt_embeds_mask,
+                        encoder_hidden_states=prompt_embeds,
+                        img_shapes=img_shapes,
+                        txt_seq_lens=prompt_embeds_mask.sum(dim=1).tolist(),
+                        joint_attention_kwargs=self.joint_attention_kwargs,
+                        return_dict=False,
+                    )[0]
+
+                if do_true_cfg:
+                    with self.transformer.cache_context("uncond"):
+                        neg_noise_pred = self.transformer(
+                            hidden_states=latents,
+                            timestep=timestep / 1000,
+                            guidance=guidance,
+                            encoder_hidden_states_mask=negative_prompt_embeds_mask,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            img_shapes=img_shapes,
+                            txt_seq_lens=negative_prompt_embeds_mask.sum(dim=1).tolist(),
+                            joint_attention_kwargs=self.joint_attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                    comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
+
+                    cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
+                    noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
+                    noise_pred = comb_pred * (cond_norm / noise_norm)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            image = self.vae.decode(latents, return_dict=False)[0][:, :, 0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return QwenImagePipelineOutput(images=image)

src/diffusers/utils/dummy_pt_objects.py

@@ -423,6 +423,21 @@ class AutoencoderKLMochi(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class AutoencoderKLQwenImage(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class AutoencoderKLTemporalDecoder(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -1038,6 +1053,21 @@ class PriorTransformer(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class QwenImageTransformer2DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class SanaControlNetModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1742,6 +1742,21 @@ class PixArtSigmaPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class QwenImagePipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class ReduxImageEncoder(metaclass=DummyObject):
     _backends = ["torch", "transformers"]